On Sun, Dec 01, 2024 at 11:54:45AM +0100, Lukas Wunner wrote: > On Sun, Dec 01, 2024 at 01:51:08PM +0530, Manivannan Sadhasivam wrote: > > The idea of pci_pwrctrl_unregister() is to remove the pwrctl device when the > > associated PCI device gets removed. When this happens, the pwrctl driver will > > turn off the power to the corresponding PCI device > > After pci_pwrctrl_unregister() is called from pci_stop_dev(), > the device may be accessed by one of the calls in pci_destroy_dev(). > > E.g. pci_doe_destroy() may set the DOE Abort bit in the DOE Control > Register if a DOE exchange is ongoing. One cannot assume that no > such exchange is ongoing merely because the device was unbound from > its driver. The PCI core may have legitimate reasons to perform a DOE > exchange or generally access config space even after the device has been > unbound. And IIUC, those accesses will fail if pwrctrl has powered the > device down. > > Another example is pcie_aspm_exit_link_state(), which will adjust ASPM > settings on device removal. > > So it seems to me that the call to pci_pwrctrl_unregister() needs to be > moved to pci_doe_destroy(). However I'm worried that will break the > symmetry with pci_pwrctrl_create_devices(), which is only called in the > !pci_dev_is_added() case. > So I took a stab at both of your suggestions: 1. Moving the pwrctrl devices creation to pci_scan_device() 2. Moving the pwrctrl devices unregister to pci_destroy_dev() Both seems to be working fine and it also allows creating pwrctrl devices for the PCI bridges without any change (for which I was planning to add one more patch earlier). But the devlink creation still needs to happen in pci_bus_add_device(), which I think is fine. Will post the patches tomorrow, thanks! - Mani -- மணிவண்ணன் சதாசிவம்
